Ruslan

Gating in place: unit, mutationm functional, still need to finish e2e testing setup.

Today this flow remains - any feature is built through Github issues, issues created from Claude mobile app (brainstormed and broken into sub tasks) or from Github mobile app. Supabase and Flutter code goes to prod from literally one click (iOS Shortcuts).

My setup to build mobile app grew substantially with access to Fable. Opus and Sonnet were definitely not capable to build and understand logical workflows' connections. Almost 0 manual involvement to add any new CI jobs. Here was the rough idea - lnkd.in/gHGubWYS

@HaoStudioCC Yeah agree. I usually have iOS text replacement to reuse prompt , then at some point (when it becomes repetitive) move it to CI. But when you can over- engineer everything with one prompt it’s so tempting to automate for the sake of automation!!!

@ikimruslan depends how much manual work you want to drop. for one-off releases cc slash commands are faster than wiring up a CI job. if youre doing this weekly then yes automate. i made a visual of the 92 cc commands so i can find what each one does when i forget: vexilo.app

Repeating question in my head when I’m trying to setup my Mobile App development pipeline with AI was always should I automate with CI or just ask Claude Code do that thing. This time it was marrying release job and release notes. I already have 2 separate jobs to write notes

I would always choose latter. Do you have similar dilemma from time to time? Or am I overthinking?

almost 100% sure nothing will be lost in between, I can easily plug in release notes reviews (from other models). And these 2 options today have the same price in terms of setting it up/building. But I think

Caught myself yesterday prompting Gemini to create script that I planned to run periodically to collect some data manually but then naturally proceeded to ask Gemini run it. Now trying to decide how I leave this in my neural net as Laziness and degradation or Efficient behavior?

Aaaand it’s back (Fable 5) , and the first thing we all have to ask it to do is:

You’ve been putting off your e2e tests cause need to deliver feature code. Cutting corners on design doc to leave time for coding. Getting Code coverage to 90% is waste of time. Guess what now it’s more important cause coding time is won back. WDYT?

Cost of writing it is low, so it can be rewritten. Verification should become the most valuable part: against spec, against guidelines, constraints, against environment. As engineers we freed up our time from writing code to write proper verification pipelines and processes.

As soon as your team really starts to adopt AI to write code you hit the wall of “more code than eye ball time” - engineers are overwhelmed with generated code. Code reviews started to be unfair, cause it’s generated. IMO code is “disposable” asset now.

Do you even have raster images in your projects if LLMs can create vector images or draw with “Canvas” and animate it? #vibecoding #AI

It sounds like a prototyping before productionazing. What do you think? Were there any experiments like this?

Got me thinking would there be any benefit in one shooting it waving all limitations even hitting context window at some point, then probably get unsatisfying results. But then ask Claude to learn from output and only then subtasks it and execute.

Is there a value in oneshooting a large complex task or should it be split immediately into lots of small ones? Tried to oneshot a complex task with 3 pages PRD, my config limits Claude with 90 turns and job failed, so went asking Opus to play architect and subtask it.

Vibecoded in 1 hr of glm 5.2 and ClaudeCode from a single prompt. Calisthenics coach right in your browser with local CV. 😮 Want a link to try? I’ll deploy quickly

@NekDenis In a new world that one prompt might be a product by itself!

Explore new paths for you and suggest them on Mondays. Claude will PROMPt YOU as an expert in field and ask permission to execute. WDYT? Aren’t people ready and need to linger on micromanaging AI?

If it does why would it need your input, should not it be a “ProActive Employer” instead? Constantly researching problems that business is working with, proactively suggesting solutions (just cap amount of tokens to not be surprised with weekly bills)…

Claude Tag - new feature from Anthropic. Tag Claude in Slack to delegate work & it will execute - researching , breaking down to tasks , implement, raise PRs. I have mixed feelings, on one hand it’s very useful & close to real workflow but on other it feels like wrong direction.